OpenSAL360: Omnidirectional Video Saliency with Fixations from 2000+ Observers

500 omnidirectional videos with a mean duration of 18.1 seconds form the largest dataset of its kind. Fixations were collected from over 2000 observers, averaging more than 84 per video, while audio tracks were played during collection. The dataset, created by ANDRYHA and last updated in 2026, consists of high-resolution 3840x1920 streams sourced from YouTube.

Use Cases

• Train saliency prediction models based on the large-scale collection of human fixation data
• Analyze viewer attention patterns in omnidirectional video based on the diverse YouTube content
• Develop multimodal attention models based on the inclusion of synchronized audio tracks
• Benchmark computer vision algorithms for 360-degree video processing based on the high-resolution video streams

Strengths

• 500 videos provide a substantial sample size for model training
• Fixation data from over 2000 observers offers a robust ground truth
• High-resolution 3840x1920 streams ensure detailed visual input

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Column-level documentation is absent; field semantics must be inferred after download
• Data may reflect geographic or content bias inherent to the YouTube source

Provenance

Source

YouTube

Collection Method

Crowdsourced eye-tracking collection

Freshness

Last updated 2026-05-29 07:31:55; freshness should be verified